SAGE: Steerable Agentic Data Generation for Deep Search with Execution Feedback

Fangyuan Xu *, Rujun Han, Yanfei Chen, Zifeng Wang, I-Hung Hsu, Jun Yan, Vishy Tirumalashetty, Eunsol Choi, Tomas Pfister, Chen-Yu Lee · 2026

Read SAGE: Steerable Agentic Data Generation for Deep Search with Execution Feedback by Fangyuan Xu *, Rujun Han, Yanfei Chen, Zifeng Wang, I-Hung Hsu, Jun Yan, Vishy Tirumalashetty, Eunsol Choi, Tomas Pfister, Chen-Yu Lee (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-18202.txt.

Description

Fangyuan Xu and co-authors generate training data for deep search agents, with execution feedback in the loop. Their 2026 preprint presents SAGE. Deep search agents answer hard questions by reasoning across many documents. They could speed up information seeking a great deal. But human annotation for such training is far too costly to gather at scale. Synthetic data is the natural answer, if it can be made good enough.

SAGE is steerable, so data can be aimed at weak spots. Execution feedback checks whether a generated task and answer actually work. That check filters out junk that pure generation would pass. Better agent training data means search assistants that fail less on hard questions.

Catalog record

CREATORFangyuan Xu *, Rujun Han, Yanfei Chen, Zifeng Wang, I-Hung Hsu, Jun Yan, Vishy Tirumalashetty, Eunsol Choi, Tomas Pfister, Chen-Yu Lee
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE54 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-18202

Browse the full library · About the Tower of Babel